What changed: from assistant to operator
Until recently, using AI at a company meant someone opening a chat, pasting text and copying the answer back. The gain was real but it died with the person: it did not scale, it could not be audited and it disappeared when they went on holiday. What changed is that models now operate inside the process — they read the incoming message, classify it, query the system, fill the field, return the result and log what they did.
That changes what you buy. You no longer buy "an AI", you buy the automation of a process with a beginning, an end and an owner. And a process has metrics: how many items a day, how long each takes today, how many come back wrong. If those three answers do not exist before the project, there will be no way to prove the gain afterwards.
Where automation pays off
High-volume repetitive service. This is the most mature case in Brazil because the channel is WhatsApp and the request is standardised. Apicio, JBKR’s own product, takes orders by voice or text, assembles them and sends them formatted to the kitchen — the team stops typing and goes back to the floor. What makes it viable is not the model: it is that the order has a closed format and the customer corrects errors within seconds.
Intake triage and classification. Tickets, emails, forms and CVs that arrive uncategorised and wait for someone to read them before being routed. The AI classifies and routes; the human only reviews what came back with low confidence. Typical gain: the queue stops depending on the working hours of whoever did the triage.
Document data extraction. Invoices, contracts, reports, bills and proposals that today become typing. This is where extraction quality matters more than the model — the subject of the article on RAG and Docling.
Internal drafts. Meeting summaries, a first pass at a proposal, product descriptions, standard replies. The gain is time to first draft, not final quality: somebody still signs it off.
Where it still does not pay off
Decisions with legal or financial consequences and no human review. Approving credit, cancelling a contract, granting an off-policy discount, calculating tax. The model is wrong with exactly the same confidence it is right, and the bill arrives at month-end close or in a tax assessment.
Processes with no owner. If nobody is accountable for the outcome today, automating only spreads the blame faster. Define the owner and the metric first, automate second.
Processes that change every week. Automation needs a minimum of stability to repay its setup cost. A rule that changes with every campaign consumes more maintenance than it saves.
Headcount reduction as the stated goal. Projects sold that way tend to stall on adoption: the people who run the process are the ones who have to teach the machine, and nobody trains their own replacement. The projects that move have a capacity goal — serve twice as much with the same team — not a cutting goal.
The maths nobody does: the cost of being wrong
Before asking "is the AI accurate?", the right question is "what does it cost when it is wrong?". Two processes with the same accuracy carry opposite risks: misclassifying a ticket costs one re-route; posting a tax wrongly costs a fine and accounting rework.
The method is simple. For each process write down three numbers: monthly volume, cost of one error and cost of one human review. If the cost of error is high, the mandatory design is AI proposes and a human confirms — and the gain comes from speed, not from skipping review. If the cost of error is low and volume is high, then it is worth automating end to end with audit sampling.
Those numbers also set the quality target. In triage, 90% accuracy with 10% going to review is usually excellent. In tax document issuing, 99% is not enough, because 1% of thousands of invoices is a recurring problem.
Governance is not bureaucracy, it is what keeps the project alive
Two risks kill automation at Brazilian companies. The first is data: pasting customer information into a public AI service breaches contracts and exposes the company under the LGPD. JBKR’s rule is to use services whose terms forbid using submitted content for training, to anonymise whatever does not need to travel identified, and to log what was sent.
The second is prompt injection: content arriving from outside — a message, a PDF, a CV — carrying instructions for the model. In automation this is serious, because the model has access to systems. The protection is to treat all external content as data and never as an order, to limit what the automation is allowed to execute, and to keep a policy layer between the user and the model. That is exactly the problem AironCore, JBKR’s other product, exists to solve.
How JBKR rolls it out
First, measure the process as it stands today: volume, time and rework rate. Without that there is no baseline. Second, small scope: one process, one channel, one team. Third, a human in the loop from day one, with the automation flagging whatever came back with low confidence. Fourth, telemetry: every processed item is logged with input, output and decision, for auditing and for improvement. Fifth, only then scale — and scaling is decided by the number, not by the feeling.
What we do not do: train a model from scratch when good process design solves it, and promise full autonomy on a process with financial consequences.
Where to start
Pick a process with daily volume, a rule you can explain in one paragraph, and an error you can undo the same day. That is the right candidate. Measure it for two weeks before automating anything.
At JBKR the initial 30-minute assessment is free and exists for exactly this triage: it produces a list of processes ranked by return and risk, with an hours estimate. Automating the wrong process is the most expensive way to start.
